Research And Implementation Of Video Scene Segmentation And Violence Detection Algorithm

With the vigorous development of information technology and multimedia technology,the number of digital videos is increasing rapidly,and new technologies are urgently needed to analyze and process them.In the process of video analysis,the video scene needs to be segmented first to extract metadata.In addition,video review is also one of the shortcomings,and a method is needed to detect violent content in the video.This paper proposes a video scene segmentation algorithm using DNN-HMM,sliding window and multi-modality.The algorithm first extracts the visual,audio and subtitle features of the shots;then uses DNN-HMM to infer the semantic elements expressed by the shots;finally uses the sliding window algorithm to segment the scenes.In addition,the early and last fusion strategies were designed for multi-modal fusion respectively.The algorithm was tested on the BBC data set,and the results showed that compared with the latest technology,it can obtain a higher average FCO score and achieve better results.This paper proposes a violent content detection algorithm.The algorithm makes full use of the multi-modal features of the video,and tries to map the input features of one modality to the high-level semantics feature of another modality,so as to realize the multi-modal fusion of high-level semantics.Experiments show that the results of this algorithm are better than the single-modal detection algorithm and the unfused detection algorithm.This paper designs a prototype system.Combining the above two algorithms can perform scene segmentation or violent content detection on uploaded videos.